AgenticDesigner

Hey, I’m Matt. Currently, I’m the Head of Design at Atlas UP. I also publish Digital Native, the newsletter for designers in the AI era.

Activity

Activity

  1. Really good write-up from @prashant_v91 on designing an app with an agent. His first screens were bad, so he added real reference apps, paper sketches for the flows he couldn't explain, and four explorations in opposite corners of two axes he picked, warmth and serenity.

    Been designing an app with Claude Code and Codex for the past few days. Writing down what worked, because my first attempts were bad. Like, really bad. Some context. I’m building a mindfulness-focussed app for myself. I started by giving Claude and Codex a broad direction and both of them messed up massively. That was on me. I hadn’t given them a single reference, so they made what every LLM makes when it has nothing to go on. 1. Read @lennysan's article on designing with LLMs before you start (link in the first comment). The bit I keep coming back to: go broad first, lots of explorations from different angles, and only then go narrow. I read it after my first few prompts. Should’ve been before. 2. Give it references. I used Appllama. Its paid but inexpensive and it lets the agent pull up real screens from real apps. Then I spent a while with the agent in Codex just talking about the apps I wanted to take from. I borrowed heavily from 3 or 4 apps. One of them was Focus Traveller. It has nothing to do with what I’m building, but I really, really like its onboarding. 3. Draw. I started with a few rough wireframes on my laptop and that was fine. When I couldn’t explain a flow in words, I drew it on paper and attached the photo (first pic). 4. Have opinions and say them out loud. “Make it better” and “improve this” do nothing for a model. I knew I wanted it mostly dark, since I’ll open it in the second half of the day and didn’t want white light in my face. So black and blue as the base, and I needed something that would stand out against that, which is how gold came in. Every time I got more specific, it got closer, and I think it actually listens more when you sound sure. 5. Second pic is what I got from the brief and the wireframe, with no skills installed. Not bad at all tbh. The dark base, the serif numbers, the question, the Elsewhere to Here scale, “A few words, if you like”. All of it is already there, and all of it came from the brief. 6. Then I installed design skills: @emilkowalski (GOAT), apple-design, swiftui-design-principles, writing-for-interfaces, etc. Third pic is the same brief after the skills. I didn’t expect the difference to be this big. Other explorations in the comments. 7. I asked for four explorations at opposite corners and gave it the axes myself, warmth on one and serenity on the other. Picked Lantern Valley and asked for the lantern to become a little character. Current home screen attached last. 8. Fable and Opus 5.5 were great at the thinking part: how the components should evolve, the design language, what the app is really about. The actual screens came out better on Codex, for whatever reason. That’s just my experience on one app, so take it with salt. 9. Somewhere along the way I started thinking of the LLM as a median designer on the team. When I could describe exactly what I wanted, it more or less got there. When all I gave it was the problem, it made something generic. Which is fine on day one when you don’t know what you want yet. That’s Lenny’s first step, go crazy there. Still iterating. Will post the onboarding once it’s done. @prashant_v91

    XToday

Shop

Shop

View All

Playground

Playground

View All